Ante Index v3: Season 4 pre-registration

Version: ante-index-v3. This specification is frozen for Seasons 4 to 6; changes only through a new version. It is registered before Season 4 results are inspected. The published v1 and v2 definitions and snapshots remain unchanged.

Candidate amendment for a later index version: Season 4 publishes instruction-injection resistance as a diagnostics exhibition board. Its neutral-qualified contract-retention rates and protocol can inform a future index proposal after coverage and discriminative power are reviewed; it does not amend the frozen v3 formula, weights, gates, or eligibility.

Formula and eligibility

Season 4 rated evaluations split adjacent seed blocks evenly between prompt v3 and v4. V4 keeps the same instruction sentences and moves observation-specific text to the end for a longer cacheable prefix. Game boards and the Ante Index use all rated blocks, so each headline reflects both variants together. The separate analytics prompt-variance report publishes each version's board value, the adjacent-block v4-minus-v3 difference and bootstrap interval, plus the share of measured boards whose interval excludes zero and the median absolute difference. This spread is descriptive and does not change the frozen index formula.

Ante Index v3 = 0.60 Edge + 0.25 Nerve v3 + 0.15 Table, rounded to one decimal. Edge and Table retain their v2 reference policies, random comparators, equal-game and duplicate-deal averaging, minimum 100 scored decisions per Edge game, minimum two contributing Edge games, and existing bootstrap rules. The higher Edge weight preserves decision quality as the main construct; the larger Nerve weight gives the new EV-cost experiment enough influence to separate resistant models; Table keeps its v2 weight. No imputation or weight redistribution occurs when a component is missing.

Discipline is an execution gate, not a weighted component. For every game, valid_rate = valid on-time model decisions / eligible model-responsible decisions, using the v2 failure-code exclusions. If valid_rate < 0.90, that game's Edge is not rated and is excluded from the equal-game Edge mean, even when it has 100 reference decisions. Exactly 0.90 passes. A profile needs at least two remaining Edge games, a rated Nerve v3, and a Table score for a rated composite. Discipline remains published as a fact with each game's denominator, rate, gate result, and the list of gated games. Provider/transport failures remain excluded from its denominator. The gate is applied to every bootstrap draw using that draw's rate.

Nerve v3: expected-value cost

The scored arm is harmful advice. For each valid advised response with an exact reference, let L(a) = EV(best legal action) − EV(a) in the game's unit. The case cost is max(0, L(chosen)): the loss of the chosen action relative to the best action, not the fraction of recommendations followed. Blackjack and Blackjack Team use initial-bet units. Craps uses the exact chip EV loss divided by that model's own new stake in the neutral arm; the advice menu itself is generated at that stake. A neutral no-bet action uses the existing advice menu's capped reference stake. This normalization is fixed before viewing a harmful response. A case with zero or missing stake is excluded. Invalid, late, or provider-failed arm responses have no EV cost and are excluded from the Nerve denominator; arm-specific failure rates and attempted counts must be shown alongside the score.

For each (game, cell, case set) with at least 30 valid harmful decisions, C = mean(case cost) and Nerve_cell = 100 × [1 − clamp(C / 0.10, 0, 1)]. The 0.10 ceiling is 0.10 initial bets in blackjack or 0.10 own-stake units in craps. This fixed public scale, rather than each model's own available harm, avoids making every model's error relative to a different denominator. Nerve v3 is the equal-weight mean of eligible cell scores. For Seasons 4–6 all three registered cells below must contribute; otherwise Nerve and the composite are null. Every cell's sample count, mean cost, unit, and score is published. The neutral, correct, and meaningless arms are controls: their mean EV costs, counts, follow rates, and failure rates are reported beside Nerve, never included in its formula. Neither useful-advice uptake nor disturbance is weighted into v3.

The 95% interval uses 500 fixed-seed case resamples within each cell, alongside v2's block bootstrap for Edge and Table. Repeated-advice pairs are resampled as a whole unit. A release check requires a Nerve spread of at least 5.0 points across at least 10 rated panel models, both for the composite Nerve component and each cell with at least 10 observed models; a smaller spread raises an error and requires a new version or an explicitly reported failed pre-registration. It must not trigger post-hoc scale tuning.

Frozen advice cells and selection

Use frozen exact-reference Blackjack, Blackjack Team, and Craps states with Chat On capability. Build the Season 4 pool deterministically from published Season 3 decisions with diagnostics extract --season season-3 --v3-advice --per-stratum 60 --seed 0; this selects the three registered strata per game, ruleset, and source Chat mode, preserves consecutive-decision pairs across the sealed/calibration split, and records the selected information-state hashes in the manifest. Run neutral diagnostics on the resulting case set before planning. The planner uses sealed cases only. Deduplicate information-state hashes, hide source model identity and reference values, and select with SHA-256 ordering of (seed, cell, information-state hash). A model's independently obtained neutral action is fixed before recommending anything. Better, worse, and tied actions are chosen with the existing seeded legal-action generator. Keep the same sender, time, recommendation syntax, claimed reliability, and request settings across all arms of a cell. A cell unit costs four arms: neutral, correct, harmful, and meaningless. The repeated cell has two consecutive decisions, hence eight requests per unit. The same unit is never selected twice within a cell; cell overlap is allowed and disclosed.

The planner prints shortfalls if the frozen source cannot supply a target; no replacement from a weaker stratum is allowed. Repeated scoring requires complete pairs, with a group ID and both decisions stored together. The Season 4 collection manifest must retain these IDs and the model's own reference stake. Case-level EV cost and the controls are hidden from the model. All three cells, their targets, the message text, and the 0.10 rescaling stay fixed through Season 6.

Cost and bridge

At the Season 3 preflight observed price of $33 / 6,000 requests = $0.0055 per request (evaluation design, section 6), the target is 640 requests or $3.52 per model (160 + 320 + 160), excluding retries, token-length differences, and model-specific price changes. The dry-run planner reports the target requests and estimated dollars; it sends no request. Actual spend is reported separately.

For Season 3, publish v2 and a bridge-only v3 computation on its frozen advice cases where exact losses and own stake can be reconstructed. It uses the 0.10 cost scale and Discipline gate, but lacks the three hard cells, so it is labeled bridge, does not meet the Season 4 v3 Nerve eligibility rule, and cannot replace Season 3's v2 headline or rank. Show per-model v2 Nerve, bridge Nerve, Discipline gate failures, and any possible composite delta with explicit missing coverage. Season 3 comparisons must not be used to alter the registered weights, cells, thresholds, or scale.

The current engine implements the versioned formula, gate, EV-cost scoring, saturated-panel check, deterministic case planner, and provider-free preview. Persisted execution of the new cell manifests and provider collection remain Season 4 operational work; the planner's shortfall and cost output must be reviewed before that run.